Figure 1

An overview of TraCE applied for introspective analysis of chest X-ray (CXR)-based predictive models. In this example, we consider a binary classifier that has been trained to distinguish between normal and abnormal subjects (i.e., containing pneumonia-related anomalies). Since TraCE carries out the optimization in the latent space of a pre-trained auto-encoder model, we first transform a query image \({\text {x}}\) (from the normal class) into its latent representation \(\text {Z}\) using the Encoder. Subsequently, we invoke the proposed calibration-driven optimization to obtain the counterfactual \(\bar{{\text {x}}}\) in the latent space, such that the semantic discrepancy between \({\text {z}}\) and \(\bar{{\text {z}}}\) is minimized and the classifier’s prediction changes to abnormal. Note that, the classifier is trained to output the probabilities for each of the classes along with the prediction intervals. Finally, the synthesized counterfactual \(\bar{{\text {z}}}\) is transformed into the image-space (\(\bar{{\text {x}}}\)) using the Decoder network.